lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001, 2002 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_dlm.h>
  34 #include <linux/init.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/random.h>
  37 #include <linux/locks.h>
  38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  39 #include <linux/buffer_head.h>
  40 #endif
  41 #include <linux/obd_lov.h>
  42 #include <linux/lprocfs_status.h>
  43
  44 static kmem_cache_t *mds_file_cache;
  45
  46 extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
  47                            obd_uuid_t *uuidarray);
  48 extern int mds_get_lovdesc(struct mds_obd  *obd, struct lov_desc *desc);
  49 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  50                                 struct ptlrpc_request *req);
  51 static int mds_cleanup(struct obd_device * obddev);
  52
  53 extern struct lprocfs_vars status_var_nm_1[];
  54 extern struct lprocfs_vars status_class_var[];
  55
  56 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  57 {
  58         return &req->rq_export->exp_obd->u.mds;
  59 }
  60
  61 static int mds_bulk_timeout(void *data)
  62 {
  63         struct ptlrpc_bulk_desc *desc = data;
  64
  65         ENTRY;
  66         CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
  67         RETURN(1);
  68 }
  69
  70 /* Assumes caller has already pushed into the kernel filesystem context */
  71 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  72                         __u64 offset)
  73 {
  74         int rc = 0;
  75         struct mds_obd *mds = mds_req2mds(req);
  76         struct ptlrpc_bulk_desc *desc;
  77         struct ptlrpc_bulk_page *bulk;
  78         struct l_wait_info lwi;
  79         char *buf;
  80         ENTRY;
  81
  82         desc = ptlrpc_prep_bulk(req->rq_connection);
  83         if (desc == NULL)
  84                 GOTO(out, rc = -ENOMEM);
  85
  86         bulk = ptlrpc_prep_bulk_page(desc);
  87         if (bulk == NULL)
  88                 GOTO(cleanup_bulk, rc = -ENOMEM);
  89
  90         OBD_ALLOC(buf, PAGE_SIZE);
  91         if (buf == NULL)
  92                 GOTO(cleanup_bulk, rc = -ENOMEM);
  93
  94         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  95
  96         if (rc != PAGE_SIZE)
  97                 GOTO(cleanup_buf, rc = -EIO);
  98
  99         bulk->bp_xid = req->rq_xid;
 100         bulk->bp_buf = buf;
 101         bulk->bp_buflen = PAGE_SIZE;
 102         desc->bd_portal = MDS_BULK_PORTAL;
 103
 104         rc = ptlrpc_send_bulk(desc);
 105         if (rc)
 106                 GOTO(cleanup_buf, rc);
 107
 108         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 109                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 110                        OBD_FAIL_MDS_SENDPAGE, rc);
 111                 ptlrpc_abort_bulk(desc);
 112                 GOTO(cleanup_buf, rc);
 113         }
 114
 115         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 116         rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
 117         if (rc) {
 118                 if (rc != -ETIMEDOUT)
 119                         LBUG();
 120                 GOTO(cleanup_buf, rc);
 121         }
 122
 123         EXIT;
 124  cleanup_buf:
 125         OBD_FREE(buf, PAGE_SIZE);
 126  cleanup_bulk:
 127         ptlrpc_free_bulk(desc);
 128  out:
 129         return rc;
 130 }
 131
 132 /*
 133  * Look up a named entry in a directory, and get an LDLM lock on it.
 134  * 'dir' is a inode for which an LDLM lock has already been taken.
 135  *
 136  * If we do not need an exclusive or write lock on this entry (e.g.
 137  * a read lock for attribute lookup only) then we do not hold the
 138  * directory on return.  It is up to the caller to know what type
 139  * of lock it is getting, and clean up appropriately.
 140  */
 141 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 142                                       struct dentry *dir, struct vfsmount **mnt,
 143                                       char *name, int namelen, int lock_mode,
 144                                       struct lustre_handle *lockh,
 145                                       int dir_lock_mode)
 146 {
 147         struct dentry *dchild;
 148         int flags = 0, rc;
 149         __u64 res_id[3] = {0};
 150         ENTRY;
 151
 152         down(&dir->d_inode->i_sem);
 153         dchild = lookup_one_len(name, dir, namelen);
 154         if (IS_ERR(dchild)) {
 155                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 156                 up(&dir->d_inode->i_sem);
 157                 LBUG();
 158                 RETURN(dchild);
 159         }
 160         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 161                 up(&dir->d_inode->i_sem);
 162                 ldlm_lock_decref(lockh, dir_lock_mode);
 163         }
 164
 165         if (lock_mode == 0 || !dchild->d_inode)
 166                 RETURN(dchild);
 167
 168         res_id[0] = dchild->d_inode->i_ino;
 169         res_id[1] = dchild->d_inode->i_generation;
 170         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 171                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 172                                    &flags, ldlm_completion_ast,
 173                                    mds_blocking_ast, NULL, 0, lockh);
 174         if (rc != ELDLM_OK) {
 175                 l_dput(dchild);
 176                 up(&dir->d_inode->i_sem);
 177                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 178         }
 179
 180         RETURN(dchild);
 181 }
 182
 183 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 184                                      struct vfsmount **mnt, int lock_mode,
 185                                      struct lustre_handle *lockh)
 186 {
 187         struct mds_obd *mds = &obd->u.mds;
 188         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 189         int flags = 0, rc;
 190         __u64 res_id[3] = {0};
 191         ENTRY;
 192
 193         if (IS_ERR(de))
 194                 RETURN(de);
 195
 196         res_id[0] = de->d_inode->i_ino;
 197         res_id[1] = de->d_inode->i_generation;
 198         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 199                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 200                                    &flags, ldlm_completion_ast,
 201                                    mds_blocking_ast, NULL, 0, lockh);
 202         if (rc != ELDLM_OK) {
 203                 l_dput(de);
 204                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 205         }
 206
 207         RETURN(retval);
 208 }
 209
 210 #ifndef DCACHE_DISCONNECTED
 211 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 212 #endif
 213
 214 /* Look up an entry by inode number. */
 215 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 216                               struct vfsmount **mnt)
 217 {
 218         /* stolen from NFS */
 219         struct super_block *sb = mds->mds_sb;
 220         unsigned long ino = fid->id;
 221         __u32 generation = fid->generation;
 222         struct inode *inode;
 223         struct list_head *lp;
 224         struct dentry *result;
 225
 226         if (ino == 0)
 227                 RETURN(ERR_PTR(-ESTALE));
 228
 229         inode = iget(sb, ino);
 230         if (inode == NULL)
 231                 RETURN(ERR_PTR(-ENOMEM));
 232
 233         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 234
 235         if (is_bad_inode(inode) ||
 236             (generation && inode->i_generation != generation)) {
 237                 /* we didn't find the right inode.. */
 238                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 239                        inode->i_ino, inode->i_nlink,
 240                        atomic_read(&inode->i_count), inode->i_generation,
 241                        generation);
 242                 iput(inode);
 243                 RETURN(ERR_PTR(-ENOENT));
 244         }
 245
 246         /* now to find a dentry. If possible, get a well-connected one */
 247         if (mnt)
 248                 *mnt = mds->mds_vfsmnt;
 249         spin_lock(&dcache_lock);
 250         list_for_each(lp, &inode->i_dentry) {
 251                 result = list_entry(lp, struct dentry, d_alias);
 252                 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
 253                         dget_locked(result);
 254                         result->d_vfs_flags |= DCACHE_REFERENCED;
 255                         spin_unlock(&dcache_lock);
 256                         iput(inode);
 257                         if (mnt)
 258                                 mntget(*mnt);
 259                         return result;
 260                 }
 261         }
 262         spin_unlock(&dcache_lock);
 263         result = d_alloc_root(inode);
 264         if (result == NULL) {
 265                 iput(inode);
 266                 return ERR_PTR(-ENOMEM);
 267         }
 268         if (mnt)
 269                 mntget(*mnt);
 270         result->d_flags |= DCACHE_DISCONNECTED;
 271         return result;
 272 }
 273
 274 /* Establish a connection to the MDS.
 275  *
 276  * This will set up an export structure for the client to hold state data
 277  * about that client, like open files, the last operation number it did
 278  * on the server, etc.
 279  */
 280 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 281                        obd_uuid_t cluuid, struct recovd_obd *recovd,
 282                        ptlrpc_recovery_cb_t recover)
 283 {
 284         struct obd_export *exp;
 285         struct mds_export_data *med;
 286         struct mds_client_data *mcd;
 287         struct list_head *p;
 288         int rc;
 289         ENTRY;
 290
 291         if (!conn || !obd || !cluuid)
 292                 RETURN(-EINVAL);
 293
 294         MOD_INC_USE_COUNT;
 295
 296         spin_lock(&obd->obd_dev_lock);
 297         list_for_each(p, &obd->obd_exports) {
 298                 exp = list_entry(p, struct obd_export, exp_obd_chain);
 299                 mcd = exp->exp_mds_data.med_mcd;
 300                 if (!mcd) {
 301                         CERROR("FYI: NULL mcd - simultaneous connects\n");
 302                         continue;
 303                 }
 304                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 305                         LASSERT(exp->exp_obd == obd);
 306
 307                         if (!list_empty(&exp->exp_conn_chain)) {
 308                                 CERROR("existing uuid/export, list not empty!\n");
 309                                 spin_unlock(&obd->obd_dev_lock);
 310                                 MOD_DEC_USE_COUNT;
 311                                 RETURN(-EALREADY);
 312                         }
 313                         conn->addr = (__u64) (unsigned long)exp;
 314                         conn->cookie = exp->exp_cookie;
 315                         spin_unlock(&obd->obd_dev_lock);
 316                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 317                                cluuid, exp);
 318                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 319                                (long long)conn->addr, (long long)conn->cookie);
 320                         MOD_DEC_USE_COUNT;
 321                         RETURN(0);
 322                 }
 323         }
 324         spin_unlock(&obd->obd_dev_lock);
 325         /* XXX There is a small race between checking the list and adding a
 326          * new connection for the same UUID, but the real threat (list
 327          * corruption when multiple different clients connect) is solved.
 328          *
 329          * There is a second race between adding the export to the list,
 330          * and filling in the client data below.  Hence skipping the case
 331          * of NULL mcd above.  We should already be controlling multiple
 332          * connects at the client, and we can't hold the spinlock over
 333          * memory allocations without risk of deadlocking.
 334          */
 335         rc = class_connect(conn, obd, cluuid);
 336         if (rc)
 337                 GOTO(out_dec, rc);
 338         exp = class_conn2export(conn);
 339         LASSERT(exp);
 340         med = &exp->exp_mds_data;
 341
 342         OBD_ALLOC(mcd, sizeof(*mcd));
 343         if (!mcd) {
 344                 CERROR("mds: out of memory for client data\n");
 345                 GOTO(out_export, rc = -ENOMEM);
 346         }
 347
 348         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 349         med->med_mcd = mcd;
 350
 351         INIT_LIST_HEAD(&med->med_open_head);
 352         spin_lock_init(&med->med_open_lock);
 353
 354         rc = mds_client_add(med, -1);
 355         if (rc)
 356                 GOTO(out_mcd, rc);
 357
 358         RETURN(0);
 359
 360 out_mcd:
 361         OBD_FREE(mcd, sizeof(*mcd));
 362 out_export:
 363         class_disconnect(conn);
 364 out_dec:
 365         MOD_DEC_USE_COUNT;
 366
 367         return rc;
 368 }
 369
 370 /* Call with med->med_open_lock held, please. */
 371 inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
 372 {
 373         struct file *file = mfd->mfd_file;
 374         LASSERT(file->private_data == mfd);
 375
 376         list_del(&mfd->mfd_list);
 377         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 378         kmem_cache_free(mds_file_cache, mfd);
 379
 380         return filp_close(file, 0);
 381 }
 382
 383 static int mds_disconnect(struct lustre_handle *conn)
 384 {
 385         struct obd_export *export = class_conn2export(conn);
 386         struct list_head *tmp, *n;
 387         struct mds_export_data *med = &export->exp_mds_data;
 388         int rc;
 389         ENTRY;
 390
 391         /*
 392          * Close any open files.
 393          */
 394         spin_lock(&med->med_open_lock);
 395         list_for_each_safe(tmp, n, &med->med_open_head) {
 396                 struct mds_file_data *mfd =
 397                         list_entry(tmp, struct mds_file_data, mfd_list);
 398                 rc = mds_close_mfd(mfd, med);
 399                 if (rc) {
 400                         /* XXX better diagnostics, with file path and stuff */
 401                         CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
 402                 }
 403         }
 404         spin_unlock(&med->med_open_lock);
 405
 406         ldlm_cancel_locks_for_export(export);
 407         mds_client_free(export);
 408
 409         rc = class_disconnect(conn);
 410         if (!rc)
 411                 MOD_DEC_USE_COUNT;
 412
 413         RETURN(rc);
 414 }
 415
 416 /*
 417  * XXX This is NOT guaranteed to flush all transactions to disk (even though
 418  *     it is equivalent to calling sync()) because it only _starts_ the flush
 419  *     and does not wait for completion.  It's better than nothing though.
 420  *     What we really want is a mild form of fsync_dev_lockfs(), but it is
 421  *     non-standard, or enabling do_sync_supers in ext3, just for this call.
 422  */
 423 static void mds_fsync_super(struct super_block *sb)
 424 {
 425         lock_kernel();
 426         lock_super(sb);
 427         if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
 428                 sb->s_op->write_super(sb);
 429         unlock_super(sb);
 430         unlock_kernel();
 431 }
 432
 433 static int mds_getstatus(struct ptlrpc_request *req)
 434 {
 435         struct mds_obd *mds = mds_req2mds(req);
 436         struct mds_body *body;
 437         int rc, size = sizeof(*body);
 438         ENTRY;
 439
 440         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 441         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 442                 CERROR("mds: out of memory for message: size=%d\n", size);
 443                 req->rq_status = -ENOMEM;
 444                 RETURN(0);
 445         }
 446
 447         /* Flush any outstanding transactions to disk so the client will
 448          * get the latest last_committed value and can drop their local
 449          * requests if they have any.  This would be fsync_super() if it
 450          * was exported.
 451          */
 452         mds_fsync_super(mds->mds_sb);
 453
 454         body = lustre_msg_buf(req->rq_repmsg, 0);
 455         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 456
 457         /* the last_committed and last_xid fields are filled in for all
 458          * replies already - no need to do so here also.
 459          */
 460         RETURN(0);
 461 }
 462
 463 static int mds_getlovinfo(struct ptlrpc_request *req)
 464 {
 465         struct mds_obd *mds = mds_req2mds(req);
 466         struct mds_status_req *streq;
 467         struct lov_desc *desc;
 468         int tgt_count;
 469         int rc, size[2] = {sizeof(*desc)};
 470         ENTRY;
 471
 472         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 473         streq->flags = NTOH__u32(streq->flags);
 474         streq->repbuf = NTOH__u32(streq->repbuf);
 475         size[1] = streq->repbuf;
 476
 477         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 478         if (rc) {
 479                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 480                 req->rq_status = -ENOMEM;
 481                 RETURN(0);
 482         }
 483
 484         desc = lustre_msg_buf(req->rq_repmsg, 0);
 485         rc = mds_get_lovdesc(mds, desc);
 486         if (rc) {
 487                 CERROR("mds_get_lovdesc error %d", rc);
 488                 req->rq_status = rc;
 489                 RETURN(0);
 490         }
 491
 492         tgt_count = le32_to_cpu(desc->ld_tgt_count);
 493         if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
 494                 CERROR("too many targets, enlarge client buffers\n");
 495                 req->rq_status = -ENOSPC;
 496                 RETURN(0);
 497         }
 498
 499         /* XXX the MDS should not really know about this */
 500         mds->mds_max_mdsize = lov_mds_md_size(tgt_count);
 501         rc = mds_get_lovtgts(mds, tgt_count,
 502                              lustre_msg_buf(req->rq_repmsg, 1));
 503         if (rc) {
 504                 CERROR("get_lovtgts error %d\n", rc);
 505                 req->rq_status = rc;
 506                 RETURN(0);
 507         }
 508         RETURN(0);
 509 }
 510
 511 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 512                      void *data, __u32 data_len, int flag)
 513 {
 514         int do_ast;
 515         ENTRY;
 516
 517         if (flag == LDLM_CB_CANCELING) {
 518                 /* Don't need to do anything here. */
 519                 RETURN(0);
 520         }
 521
 522         /* XXX layering violation!  -phil */
 523         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 524         lock->l_flags |= LDLM_FL_CBPENDING;
 525         do_ast = (!lock->l_readers && !lock->l_writers);
 526         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 527
 528         if (do_ast) {
 529                 struct lustre_handle lockh;
 530                 int rc;
 531
 532                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 533                 ldlm_lock2handle(lock, &lockh);
 534                 rc = ldlm_cli_cancel(&lockh);
 535                 if (rc < 0)
 536                         CERROR("ldlm_cli_cancel: %d\n", rc);
 537         } else
 538                 LDLM_DEBUG(lock, "Lock still has references, will be"
 539                            "cancelled later");
 540         RETURN(0);
 541 }
 542
 543 int mds_pack_md(struct mds_obd *mds, struct ptlrpc_request *req,
 544                 int offset, struct mds_body *body, struct inode *inode)
 545 {
 546         struct lov_mds_md *lmm;
 547         int lmm_size = req->rq_repmsg->buflens[offset];
 548         int rc;
 549
 550         if (lmm_size == 0) {
 551                 CDEBUG(D_INFO, "no space reserved for inode %u MD\n", inode->i_ino);
 552                 RETURN(0);
 553         }
 554
 555         lmm = lustre_msg_buf(req->rq_repmsg, offset);
 556
 557         /* I don't really like this, but it is a sanity check on the client
 558          * MD request.  However, if the client doesn't know how much space
 559          * to reserve for the MD, this shouldn't be fatal either...
 560          */
 561         if (lmm_size > mds->mds_max_mdsize) {
 562                 CERROR("Reading MD for inode %u of %d bytes > max %d\n",
 563                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 564                 // RETURN(-EINVAL);
 565         }
 566
 567         /* We don't need to store the reply size, because this buffer is
 568          * discarded right after unpacking, and the LOV can figure out the
 569          * size itself from the ost count.
 570          */
 571         if ((rc = mds_fs_get_md(mds, inode, lmm, lmm_size)) < 0) {
 572                 CDEBUG(D_INFO, "No md for ino %u: rc = %d\n", inode->i_ino, rc);
 573         } else if (rc > 0) {
 574                 body->valid |= OBD_MD_FLEASIZE;
 575                 rc = 0;
 576         }
 577
 578         return rc;
 579 }
 580
 581 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 582                                 struct ptlrpc_request *req,
 583                                 struct mds_body *reqbody, int reply_off)
 584 {
 585         struct mds_body *body;
 586         struct inode *inode = dentry->d_inode;
 587         int rc = 0;
 588         ENTRY;
 589
 590         if (inode == NULL)
 591                 RETURN(-ENOENT);
 592
 593         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 594
 595         mds_pack_inode2fid(&body->fid1, inode);
 596         mds_pack_inode2body(body, inode);
 597
 598         if (S_ISREG(inode->i_mode)) {
 599                 rc = mds_pack_md(mds, req, reply_off + 1, body, inode);
 600         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 601                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 602                 int len = req->rq_repmsg->buflens[reply_off + 1];
 603
 604                 rc = inode->i_op->readlink(dentry, symname, len);
 605                 if (rc < 0) {
 606                         CERROR("readlink failed: %d\n", rc);
 607                 } else {
 608                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 609                         body->valid |= OBD_MD_LINKNAME;
 610                 }
 611         }
 612         RETURN(rc);
 613 }
 614
 615 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 616 {
 617         struct mds_obd *mds = mds_req2mds(req);
 618         struct obd_device *obd = req->rq_export->exp_obd;
 619         struct obd_run_ctxt saved;
 620         struct mds_body *body;
 621         struct dentry *de = NULL, *dchild = NULL;
 622         struct inode *dir;
 623         struct lustre_handle lockh;
 624         char *name;
 625         int namelen, flags = 0, lock_mode, rc = 0;
 626         struct obd_ucred uc;
 627         __u64 res_id[3] = {0, 0, 0};
 628         ENTRY;
 629
 630         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 631
 632         if (req->rq_reqmsg->bufcount <= offset + 1) {
 633                 LBUG();
 634                 GOTO(out_pre_de, rc = -EINVAL);
 635         }
 636
 637         body = lustre_msg_buf(req->rq_reqmsg, offset);
 638         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 639         namelen = req->rq_reqmsg->buflens[offset + 1];
 640         /* requests were at offset 2, replies go back at 1 */
 641         if (offset)
 642                 offset = 1;
 643
 644         uc.ouc_fsuid = body->fsuid;
 645         uc.ouc_fsgid = body->fsgid;
 646         uc.ouc_cap = body->capability;
 647         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 648         de = mds_fid2dentry(mds, &body->fid1, NULL);
 649         if (IS_ERR(de)) {
 650                 GOTO(out_pre_de, rc = -ENOENT);
 651         }
 652
 653         dir = de->d_inode;
 654         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 655
 656         lock_mode = LCK_PR;
 657         res_id[0] = dir->i_ino;
 658         res_id[1] = dir->i_generation;
 659
 660         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 661                              NULL, 0, lock_mode, &lockh);
 662         if (rc == 0) {
 663                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 664                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 665                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 666                                       &flags, ldlm_completion_ast,
 667                                       mds_blocking_ast, NULL, 0, &lockh);
 668                 if (rc != ELDLM_OK) {
 669                         CERROR("lock enqueue: err: %d\n", rc);
 670                         GOTO(out_create_de, rc = -EIO);
 671                 }
 672         }
 673         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 674
 675         down(&dir->i_sem);
 676         dchild = lookup_one_len(name, de, namelen - 1);
 677         if (IS_ERR(dchild)) {
 678                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 679                 up(&dir->i_sem);
 680                 GOTO(out_create_dchild, rc = PTR_ERR(dchild));
 681         }
 682
 683         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 684
 685         EXIT;
 686 out_create_dchild:
 687         l_dput(dchild);
 688         up(&dir->i_sem);
 689         ldlm_lock_decref(&lockh, lock_mode);
 690 out_create_de:
 691         l_dput(de);
 692 out_pre_de:
 693         req->rq_status = rc;
 694         pop_ctxt(&saved);
 695         return 0;
 696 }
 697
 698 static int mds_getattr(int offset, struct ptlrpc_request *req)
 699 {
 700         struct mds_obd *mds = mds_req2mds(req);
 701         struct obd_run_ctxt saved;
 702         struct dentry *de;
 703         struct inode *inode;
 704         struct mds_body *body;
 705         struct obd_ucred uc;
 706         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 707         ENTRY;
 708
 709         body = lustre_msg_buf(req->rq_reqmsg, offset);
 710         uc.ouc_fsuid = body->fsuid;
 711         uc.ouc_fsgid = body->fsgid;
 712         uc.ouc_cap = body->capability;
 713         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 714         de = mds_fid2dentry(mds, &body->fid1, NULL);
 715         if (IS_ERR(de)) {
 716                 rc = req->rq_status = -ENOENT;
 717                 GOTO(out_pop, PTR_ERR(de));
 718         }
 719
 720         inode = de->d_inode;
 721         if (S_ISREG(body->fid1.f_type)) {
 722                 int rc = mds_fs_get_md(mds, inode, NULL, 0);
 723                 CDEBUG(D_INODE, "got %d bytes MD data for inode %u\n",
 724                        rc, inode->i_ino);
 725                 if (rc < 0) {
 726                         if (rc != -ENODATA)
 727                                 CERROR("error getting inode %u MD: rc = %d\n",
 728                                        inode->i_ino, rc);
 729                         size[bufcount] = 0;
 730                 } else if (rc > mds->mds_max_mdsize) {
 731                         size[bufcount] = 0;
 732                         CERROR("MD size %d larger than maximum possible %u\n",
 733                                rc, mds->mds_max_mdsize);
 734                 } else
 735                         size[bufcount] = rc;
 736                 bufcount++;
 737         } else if (body->valid & OBD_MD_LINKNAME) {
 738                 size[bufcount] = MIN(inode->i_size + 1, body->size);
 739                 bufcount++;
 740                 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
 741                        inode->i_size + 1, body->size);
 742         }
 743
 744         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 745                 CERROR("failed MDS_GETATTR_PACK test\n");
 746                 req->rq_status = -ENOMEM;
 747                 GOTO(out, rc = -ENOMEM);
 748         }
 749
 750         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 751                              &req->rq_repmsg);
 752         if (rc) {
 753                 CERROR("out of memoryK\n");
 754                 req->rq_status = rc;
 755                 GOTO(out, rc);
 756         }
 757
 758         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 759
 760 out:
 761         l_dput(de);
 762 out_pop:
 763         pop_ctxt(&saved);
 764         RETURN(rc);
 765 }
 766
 767 static int mds_statfs(struct ptlrpc_request *req)
 768 {
 769         struct mds_obd *mds = mds_req2mds(req);
 770         struct obd_statfs *osfs;
 771         struct statfs sfs;
 772         int rc, size = sizeof(*osfs);
 773         ENTRY;
 774
 775         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 776         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 777                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 778                 GOTO(out, rc);
 779         }
 780
 781         rc = mds_fs_statfs(mds, &sfs);
 782         if (rc) {
 783                 CERROR("mds: statfs failed: rc %d\n", rc);
 784                 GOTO(out, rc);
 785         }
 786         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 787         memset(osfs, 0, size);
 788         statfs_pack(osfs, &sfs);
 789         obd_statfs_pack(osfs, osfs);
 790
 791 out:
 792         req->rq_status = rc;
 793         RETURN(0);
 794 }
 795
 796 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
 797 {
 798         struct mds_file_data *mfd = NULL;
 799
 800         if (!handle || !handle->addr)
 801                 RETURN(NULL);
 802
 803         mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
 804         if (!kmem_cache_validate(mds_file_cache, mfd))
 805                 RETURN(NULL);
 806
 807         if (mfd->mfd_servercookie != handle->cookie)
 808                 RETURN(NULL);
 809
 810         return mfd;
 811 }
 812
 813 static int mds_store_md(struct mds_obd *mds, struct ptlrpc_request *req,
 814                         int offset, struct mds_body *body, struct inode *inode)
 815 {
 816         struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, offset);
 817         int lmm_size = req->rq_reqmsg->buflens[offset];
 818         struct obd_run_ctxt saved;
 819         struct obd_ucred uc;
 820         void *handle;
 821         int rc, rc2;
 822         ENTRY;
 823
 824         /* I don't really like this, but it is a sanity check on the client
 825          * MD request.
 826          */
 827         if (lmm_size > mds->mds_max_mdsize) {
 828                 CERROR("Saving MD for inode %u of %d bytes > max %d\n",
 829                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 830                 //RETURN(-EINVAL);
 831         }
 832
 833         CDEBUG(D_INODE, "storing %d bytes MD for inode %u\n",
 834                lmm_size, inode->i_ino);
 835         uc.ouc_fsuid = body->fsuid;
 836         uc.ouc_fsgid = body->fsgid;
 837         uc.ouc_cap = body->capability;
 838         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 839         handle = mds_fs_start(mds, inode, MDS_FSOP_SETATTR);
 840         if (!handle)
 841                 GOTO(out_ea, rc = -ENOMEM);
 842
 843         rc = mds_fs_set_md(mds, inode, handle, lmm, lmm_size);
 844         if (!rc)
 845                 rc = mds_update_last_rcvd(mds, handle, req);
 846
 847         rc2 = mds_fs_commit(mds, inode, handle);
 848         if (rc2 && !rc)
 849                 rc = rc2;
 850 out_ea:
 851         pop_ctxt(&saved);
 852
 853         RETURN(rc);
 854 }
 855
 856 static int mds_open(struct ptlrpc_request *req)
 857 {
 858         struct mds_obd *mds = mds_req2mds(req);
 859         struct mds_body *body;
 860         struct mds_export_data *med;
 861         struct mds_file_data *mfd;
 862         struct dentry *de;
 863         struct file *file;
 864         struct vfsmount *mnt;
 865         __u32 flags;
 866         struct list_head *tmp;
 867         int rc, size = sizeof(*body);
 868         ENTRY;
 869
 870         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 871                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 872                 req->rq_status = -ENOMEM;
 873                 RETURN(-ENOMEM);
 874         }
 875
 876         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 877         if (rc) {
 878                 CERROR("mds: pack error: rc = %d\n", rc);
 879                 req->rq_status = rc;
 880                 RETURN(rc);
 881         }
 882
 883         body = lustre_msg_buf(req->rq_reqmsg, 0);
 884
 885         /* was this animal open already and the client lost the reply? */
 886         /* XXX need some way to detect a reopen, to avoid locked list walks */
 887         med = &req->rq_export->exp_mds_data;
 888         spin_lock(&med->med_open_lock);
 889         list_for_each(tmp, &med->med_open_head) {
 890                 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
 891                 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
 892                             sizeof(mfd->mfd_clienthandle)) &&
 893                     body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
 894                         de = mfd->mfd_file->f_dentry;
 895                         spin_unlock(&med->med_open_lock);
 896                         CERROR("Re opening "LPD64"\n", body->fid1.id);
 897                         GOTO(out_pack, rc = 0);
 898                 }
 899         }
 900         spin_unlock(&med->med_open_lock);
 901
 902         mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
 903         if (!mfd) {
 904                 CERROR("mds: out of memory\n");
 905                 req->rq_status = -ENOMEM;
 906                 RETURN(0);
 907         }
 908
 909         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 910         if (IS_ERR(de))
 911                 GOTO(out_free, rc = PTR_ERR(de));
 912
 913         /* check if this inode has seen a delayed object creation */
 914         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
 915                 rc = mds_store_md(mds, req, 1, body, de->d_inode);
 916                 if (rc) {
 917                         l_dput(de);
 918                         mntput(mnt);
 919                         GOTO(out_free, rc);
 920                 }
 921         }
 922
 923         flags = body->flags;
 924         /* dentry_open does a dput(de) and mntput(mnt) on error */
 925         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 926         if (IS_ERR(file)) {
 927                 rc = PTR_ERR(file);
 928                 GOTO(out_free, 0);
 929         }
 930
 931         file->private_data = mfd;
 932         mfd->mfd_file = file;
 933         memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
 934         get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
 935         spin_lock(&med->med_open_lock);
 936         list_add(&mfd->mfd_list, &med->med_open_head);
 937         spin_unlock(&med->med_open_lock);
 938
 939 out_pack:
 940         body = lustre_msg_buf(req->rq_repmsg, 0);
 941         mds_pack_inode2fid(&body->fid1, de->d_inode);
 942         mds_pack_inode2body(body, de->d_inode);
 943         body->handle.addr = (__u64)(unsigned long)mfd;
 944         body->handle.cookie = mfd->mfd_servercookie;
 945         CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
 946                mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
 947         RETURN(0);
 948
 949 out_free:
 950         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 951         kmem_cache_free(mds_file_cache, mfd);
 952         req->rq_status = rc;
 953         RETURN(0);
 954 }
 955
 956 static int mds_close(struct ptlrpc_request *req)
 957 {
 958         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 959         struct mds_body *body;
 960         struct mds_file_data *mfd;
 961         int rc;
 962         ENTRY;
 963
 964         body = lustre_msg_buf(req->rq_reqmsg, 0);
 965
 966         mfd = mds_handle2mfd(&body->handle);
 967         if (!mfd) {
 968                 CERROR("no handle for file close "LPD64
 969                        ": addr "LPX64", cookie "LPX64"\n",
 970                        body->fid1.id, body->handle.addr, body->handle.cookie);
 971                 RETURN(-ESTALE);
 972         }
 973
 974         spin_lock(&med->med_open_lock);
 975         req->rq_status = mds_close_mfd(mfd, med);
 976         spin_unlock(&med->med_open_lock);
 977
 978         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 979                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
 980                 req->rq_status = -ENOMEM;
 981                 RETURN(-ENOMEM);
 982         }
 983
 984         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 985         if (rc) {
 986                 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
 987                 req->rq_status = rc;
 988         }
 989
 990         RETURN(0);
 991 }
 992
 993 static int mds_readpage(struct ptlrpc_request *req)
 994 {
 995         struct mds_obd *mds = mds_req2mds(req);
 996         struct vfsmount *mnt;
 997         struct dentry *de;
 998         struct file *file;
 999         struct mds_body *body, *repbody;
1000         struct obd_run_ctxt saved;
1001         int rc, size = sizeof(*body);
1002         struct obd_ucred uc;
1003         ENTRY;
1004
1005         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
1006         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1007                 CERROR("mds: out of memory\n");
1008                 GOTO(out, rc = -ENOMEM);
1009         }
1010
1011         body = lustre_msg_buf(req->rq_reqmsg, 0);
1012         uc.ouc_fsuid = body->fsuid;
1013         uc.ouc_fsgid = body->fsgid;
1014         uc.ouc_cap = body->capability;
1015         push_ctxt(&saved, &mds->mds_ctxt, &uc);
1016         de = mds_fid2dentry(mds, &body->fid1, &mnt);
1017         if (IS_ERR(de))
1018                 GOTO(out_pop, rc = PTR_ERR(de));
1019
1020         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
1021
1022         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1023         /* note: in case of an error, dentry_open puts dentry */
1024         if (IS_ERR(file))
1025                 GOTO(out_pop, rc = PTR_ERR(file));
1026
1027         repbody = lustre_msg_buf(req->rq_repmsg, 0);
1028         repbody->size = file->f_dentry->d_inode->i_size;
1029         repbody->valid = OBD_MD_FLSIZE;
1030
1031         /* to make this asynchronous make sure that the handling function
1032            doesn't send a reply when this function completes. Instead a
1033            callback function would send the reply */
1034         /* note: in case of an error, dentry_open puts dentry */
1035         rc = mds_sendpage(req, file, body->size);
1036
1037         filp_close(file, 0);
1038 out_pop:
1039         pop_ctxt(&saved);
1040 out:
1041         req->rq_status = rc;
1042         RETURN(0);
1043 }
1044
1045 int mds_reint(struct ptlrpc_request *req, int offset)
1046 {
1047         int rc;
1048         struct mds_update_record rec;
1049
1050         rc = mds_update_unpack(req, offset, &rec);
1051         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1052                 CERROR("invalid record\n");
1053                 req->rq_status = -EINVAL;
1054                 RETURN(0);
1055         }
1056         /* rc will be used to interrupt a for loop over multiple records */
1057         rc = mds_reint_rec(&rec, offset, req);
1058         return rc;
1059 }
1060
1061 int mds_handle(struct ptlrpc_request *req)
1062 {
1063         int rc;
1064         ENTRY;
1065
1066         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1067         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
1068                 CERROR("lustre_mds: Invalid request\n");
1069                 GOTO(out, rc);
1070         }
1071
1072         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
1073                 GOTO(out, rc = -ENOTCONN);
1074
1075         LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
1076
1077         switch (req->rq_reqmsg->opc) {
1078         case MDS_CONNECT:
1079                 CDEBUG(D_INODE, "connect\n");
1080                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1081                 rc = target_handle_connect(req);
1082                 break;
1083
1084         case MDS_DISCONNECT:
1085                 CDEBUG(D_INODE, "disconnect\n");
1086                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1087                 rc = target_handle_disconnect(req);
1088                 goto out;
1089
1090         case MDS_GETSTATUS:
1091                 CDEBUG(D_INODE, "getstatus\n");
1092                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1093                 rc = mds_getstatus(req);
1094                 break;
1095
1096         case MDS_GETLOVINFO:
1097                 CDEBUG(D_INODE, "getlovinfo\n");
1098                 rc = mds_getlovinfo(req);
1099                 break;
1100
1101         case MDS_GETATTR:
1102                 CDEBUG(D_INODE, "getattr\n");
1103                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1104                 rc = mds_getattr(0, req);
1105                 break;
1106
1107         case MDS_STATFS:
1108                 CDEBUG(D_INODE, "statfs\n");
1109                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1110                 rc = mds_statfs(req);
1111                 break;
1112
1113         case MDS_READPAGE:
1114                 CDEBUG(D_INODE, "readpage\n");
1115                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1116                 rc = mds_readpage(req);
1117
1118                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1119                         return 0;
1120                 break;
1121
1122         case MDS_REINT: {
1123                 int size = sizeof(struct mds_body);
1124                 CDEBUG(D_INODE, "reint\n");
1125                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1126
1127                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1128                                      &req->rq_repmsg);
1129                 if (rc) {
1130                         req->rq_status = rc;
1131                         break;
1132                 }
1133                 rc = mds_reint(req, 0);
1134                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1135                 break;
1136                 }
1137
1138         case MDS_OPEN:
1139                 CDEBUG(D_INODE, "open\n");
1140                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1141                 rc = mds_open(req);
1142                 break;
1143
1144         case MDS_CLOSE:
1145                 CDEBUG(D_INODE, "close\n");
1146                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1147                 rc = mds_close(req);
1148                 break;
1149
1150         case LDLM_ENQUEUE:
1151                 CDEBUG(D_INODE, "enqueue\n");
1152                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1153                 rc = ldlm_handle_enqueue(req);
1154                 break;
1155         case LDLM_CONVERT:
1156                 CDEBUG(D_INODE, "convert\n");
1157                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1158                 rc = ldlm_handle_convert(req);
1159                 break;
1160         case LDLM_BL_CALLBACK:
1161         case LDLM_CP_CALLBACK:
1162                 CDEBUG(D_INODE, "callback\n");
1163                 CERROR("callbacks should not happen on MDS\n");
1164                 LBUG();
1165                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1166                 break;
1167         default:
1168                 rc = ptlrpc_error(req->rq_svc, req);
1169                 RETURN(rc);
1170         }
1171
1172         EXIT;
1173
1174         if (!rc) {
1175                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1176                 struct mds_obd *mds = mds_req2mds(req);
1177
1178                 req->rq_repmsg->last_xid =
1179                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1180                 req->rq_repmsg->last_committed =
1181                         HTON__u64(mds->mds_last_committed);
1182                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1183                        (unsigned long long)mds->mds_last_rcvd,
1184                        (unsigned long long)mds->mds_last_committed,
1185                        cpu_to_le32(req->rq_xid));
1186         }
1187  out:
1188         if (rc) {
1189                 CERROR("mds: processing error (opcode %d): %d\n",
1190                        req->rq_reqmsg->opc, rc);
1191                 ptlrpc_error(req->rq_svc, req);
1192         } else {
1193                 CDEBUG(D_NET, "sending reply\n");
1194                 ptlrpc_reply(req->rq_svc, req);
1195         }
1196         return 0;
1197 }
1198
1199 /* Update the server data on disk.  This stores the new mount_count and
1200  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1201  * then the server last_rcvd value may be less than that of the clients.
1202  * This will alert us that we may need to do client recovery.
1203  *
1204  * Assumes we are already in the server filesystem context.
1205  *
1206  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1207  */
1208 static
1209 int mds_update_server_data(struct mds_obd *mds)
1210 {
1211         struct mds_server_data *msd = mds->mds_server_data;
1212         struct file *filp = mds->mds_rcvd_filp;
1213         loff_t off = 0;
1214         int rc;
1215
1216         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1217         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1218
1219         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1220                (unsigned long long)mds->mds_mount_count,
1221                (unsigned long long)mds->mds_last_rcvd);
1222         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1223         if (rc != sizeof(*msd)) {
1224                 CERROR("error writing MDS server data: rc = %d\n", rc);
1225                 if (rc > 0)
1226                         RETURN(-EIO);
1227                 RETURN(rc);
1228         }
1229 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1230         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1231 #else
1232         rc = file_fsync(filp,  filp->f_dentry, 1);
1233 #endif
1234         if (rc)
1235                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1236
1237         return 0;
1238 }
1239
1240 /* Do recovery actions for the MDS */
1241 static int mds_recover(struct obd_device *obddev)
1242 {
1243         struct mds_obd *mds = &obddev->u.mds;
1244         struct obd_run_ctxt saved;
1245         int rc;
1246
1247         /* This happens at the end when recovery is complete */
1248         ++mds->mds_mount_count;
1249         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1250         rc = mds_update_server_data(mds);
1251         pop_ctxt(&saved);
1252
1253         return rc;
1254 }
1255
1256 /* mount the file system (secretly) */
1257 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1258 {
1259         struct obd_ioctl_data* data = buf;
1260         struct mds_obd *mds = &obddev->u.mds;
1261         struct vfsmount *mnt;
1262         int rc = 0;
1263         ENTRY;
1264
1265         MOD_INC_USE_COUNT;
1266 #ifdef CONFIG_DEV_RDONLY
1267         dev_clear_rdonly(2);
1268 #endif
1269         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1270                 GOTO(err_dec, rc = -EINVAL);
1271
1272         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1273
1274         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1275         if (IS_ERR(mnt)) {
1276                 rc = PTR_ERR(mnt);
1277                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1278                 GOTO(err_kfree, rc);
1279         }
1280
1281         CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
1282         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1283         if (!mds->mds_sb)
1284                 GOTO(err_put, rc = -ENODEV);
1285
1286         spin_lock_init(&mds->mds_last_lock);
1287         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1288         rc = mds_fs_setup(obddev, mnt);
1289         if (rc) {
1290                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1291                 GOTO(err_put, rc);
1292         }
1293
1294         obddev->obd_namespace =
1295                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1296         if (obddev->obd_namespace == NULL) {
1297                 mds_cleanup(obddev);
1298                 GOTO(err_fs, rc = -ENOMEM);
1299         }
1300
1301
1302         rc = mds_recover(obddev);
1303         if (rc)
1304                 GOTO(err_fs, rc);
1305
1306         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1307                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1308
1309         RETURN(0);
1310
1311 err_fs:
1312         mds_fs_cleanup(obddev);
1313 err_put:
1314         unlock_kernel();
1315         mntput(mds->mds_vfsmnt);
1316         mds->mds_sb = 0;
1317         lock_kernel();
1318 err_kfree:
1319         kfree(mds->mds_fstype);
1320 err_dec:
1321         MOD_DEC_USE_COUNT;
1322         RETURN(rc);
1323 }
1324
1325 static int mds_cleanup(struct obd_device *obddev)
1326 {
1327         struct super_block *sb;
1328         struct mds_obd *mds = &obddev->u.mds;
1329         struct obd_run_ctxt saved;
1330         ENTRY;
1331
1332         sb = mds->mds_sb;
1333         if (!mds->mds_sb)
1334                 RETURN(0);
1335
1336         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1337         mds_update_server_data(mds);
1338
1339         if (mds->mds_rcvd_filp) {
1340                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1341                 mds->mds_rcvd_filp = NULL;
1342
1343                 if (rc)
1344                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1345         }
1346         pop_ctxt(&saved);
1347
1348         unlock_kernel();
1349         mntput(mds->mds_vfsmnt);
1350         mds->mds_sb = 0;
1351         kfree(mds->mds_fstype);
1352
1353         ldlm_namespace_free(obddev->obd_namespace);
1354
1355         lock_kernel();
1356 #ifdef CONFIG_DEV_RDONLY
1357         dev_clear_rdonly(2);
1358 #endif
1359         mds_fs_cleanup(obddev);
1360
1361         MOD_DEC_USE_COUNT;
1362         RETURN(0);
1363 }
1364
1365 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1366                               ldlm_mode_t mode, int flags, void *data)
1367 {
1368         struct ptlrpc_request *req = req_cookie;
1369         int rc = 0;
1370         ENTRY;
1371
1372         if (!req_cookie)
1373                 RETURN(0);
1374
1375         if (req->rq_reqmsg->bufcount > 1) {
1376                 /* an intent needs to be considered */
1377                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1378                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1379                 struct mds_body *mds_rep;
1380                 struct ldlm_reply *rep;
1381                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1382                 int rc, size[3] = {sizeof(struct ldlm_reply),
1383                                                   sizeof(struct mds_body),
1384                                                   mds->mds_max_mdsize};
1385
1386                 it->opc = NTOH__u64(it->opc);
1387
1388                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1389                            ldlm_it2str(it->opc));
1390
1391                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1392                                      &req->rq_repmsg);
1393                 if (rc) {
1394                         rc = req->rq_status = -ENOMEM;
1395                         RETURN(rc);
1396                 }
1397
1398                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1399                 rep->lock_policy_res1 = 1;
1400
1401                 /* execute policy */
1402                 switch ((long)it->opc) {
1403                 case IT_CREAT|IT_OPEN:
1404                         rc = mds_reint(req, 2);
1405                         if (rc || (req->rq_status != 0 &&
1406                                    req->rq_status != -EEXIST)) {
1407                                 rep->lock_policy_res2 = req->rq_status;
1408                                 RETURN(ELDLM_LOCK_ABORTED);
1409                         }
1410                         break;
1411                 case IT_CREAT:
1412                 case IT_MKDIR:
1413                 case IT_MKNOD:
1414                 case IT_RENAME2:
1415                 case IT_LINK2:
1416                 case IT_RMDIR:
1417                 case IT_SYMLINK:
1418                 case IT_UNLINK:
1419                         rc = mds_reint(req, 2);
1420                         if (rc || (req->rq_status != 0 &&
1421                                    req->rq_status != -EISDIR &&
1422                                    req->rq_status != -ENOTDIR)) {
1423                                 rep->lock_policy_res2 = req->rq_status;
1424                                 RETURN(ELDLM_LOCK_ABORTED);
1425                         }
1426                         break;
1427                 case IT_GETATTR:
1428                 case IT_LOOKUP:
1429                 case IT_OPEN:
1430                 case IT_READDIR:
1431                 case IT_READLINK:
1432                 case IT_RENAME:
1433                 case IT_LINK:
1434                 case IT_SETATTR:
1435                         rc = mds_getattr_name(2, req);
1436                         /* FIXME: we need to sit down and decide on who should
1437                          * set req->rq_status, who should return negative and
1438                          * positive return values, and what they all mean. */
1439                         if (rc || req->rq_status != 0) {
1440                                 rep->lock_policy_res2 = req->rq_status;
1441                                 RETURN(ELDLM_LOCK_ABORTED);
1442                         }
1443                         break;
1444                 case IT_READDIR|IT_OPEN:
1445                         LBUG();
1446                         break;
1447                 default:
1448                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1449                         LBUG();
1450                 }
1451
1452                 /* We don't bother returning a lock to the client for a file
1453                  * or directory we are removing.
1454                  *
1455                  * As for link and rename, there is no reason for the client
1456                  * to get a lock on the target at this point.  If they are
1457                  * going to modify the file/directory later they will get a
1458                  * lock at that time.
1459                  */
1460                 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1461                                IT_RENAME | IT_RENAME2))
1462                         RETURN(ELDLM_LOCK_ABORTED);
1463
1464                 rep->lock_policy_res2 = req->rq_status;
1465                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1466
1467                 /* If the client is about to open a file that doesn't have an MD
1468                  * stripe record, it's going to need a write lock. */
1469                 if (it->opc & IT_OPEN &&
1470                     !(lustre_msg_get_op_flags(req->rq_reqmsg)&MDS_OPEN_HAS_EA)){
1471                         LDLM_DEBUG(lock, "open with no EA; returning PW lock");
1472                         lock->l_req_mode = LCK_PW;
1473                 }
1474
1475                 if (flags & LDLM_FL_INTENT_ONLY) {
1476                         LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
1477                         RETURN(ELDLM_LOCK_ABORTED);
1478                 }
1479                 /* Give the client a lock on the child object, instead of the
1480                  * parent that it requested. */
1481                 new_resid[0] = NTOH__u32(mds_rep->ino);
1482                 new_resid[1] = NTOH__u32(mds_rep->generation);
1483                 if (new_resid[0] == 0)
1484                         LBUG();
1485                 old_res = lock->l_resource->lr_name[0];
1486
1487                 ldlm_lock_change_resource(lock, new_resid);
1488                 if (lock->l_resource == NULL) {
1489                         LBUG();
1490                         RETURN(-ENOMEM);
1491                 }
1492                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1493                            (long)old_res);
1494                 RETURN(ELDLM_LOCK_CHANGED);
1495         } else {
1496                 int size = sizeof(struct ldlm_reply);
1497                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1498                                      &req->rq_repmsg);
1499                 if (rc) {
1500                         LBUG();
1501                         RETURN(-ENOMEM);
1502                 }
1503         }
1504         RETURN(rc);
1505 }
1506
1507 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1508 {
1509         return lprocfs_reg_obd(dev, status_var_nm_1, dev);
1510 }
1511
1512 int mds_detach(struct obd_device *dev)
1513 {
1514         return lprocfs_dereg_obd(dev);
1515 }
1516
1517 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1518 {
1519         int i;
1520         //        struct obd_ioctl_data* data = buf;
1521         struct mds_obd *mds = &obddev->u.mds;
1522         int rc = 0;
1523         ENTRY;
1524
1525         MOD_INC_USE_COUNT;
1526
1527         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1528                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1529                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1530                                            "self", mds_handle, "mds");
1531         if (!mds->mds_service) {
1532                 CERROR("failed to start service\n");
1533                 GOTO(err_dec, rc = -EINVAL);
1534         }
1535
1536         for (i = 0; i < MDT_NUM_THREADS; i++) {
1537                 char name[32];
1538                 sprintf(name, "ll_mdt_%02d", i);
1539                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1540                 if (rc) {
1541                         CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
1542                         GOTO(err_thread, rc);
1543                 }
1544         }
1545
1546         RETURN(0);
1547
1548 err_thread:
1549         ptlrpc_stop_all_threads(mds->mds_service);
1550         ptlrpc_unregister_service(mds->mds_service);
1551 err_dec:
1552         MOD_DEC_USE_COUNT;
1553         RETURN(rc);
1554 }
1555
1556
1557 static int mdt_cleanup(struct obd_device *obddev)
1558 {
1559         struct mds_obd *mds = &obddev->u.mds;
1560         ENTRY;
1561
1562         ptlrpc_stop_all_threads(mds->mds_service);
1563         ptlrpc_unregister_service(mds->mds_service);
1564
1565         MOD_DEC_USE_COUNT;
1566         RETURN(0);
1567 }
1568
1569 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1570                          int len, void *karg, void *uarg);
1571
1572 /* use obd ops to offer management infrastructure */
1573 static struct obd_ops mds_obd_ops = {
1574         o_attach:      mds_attach,
1575         o_detach:      mds_detach,
1576         o_connect:     mds_connect,
1577         o_disconnect:  mds_disconnect,
1578         o_setup:       mds_setup,
1579         o_cleanup:     mds_cleanup,
1580         o_iocontrol:   mds_iocontrol
1581 };
1582
1583 static struct obd_ops mdt_obd_ops = {
1584         o_setup:       mdt_setup,
1585         o_cleanup:     mdt_cleanup,
1586 };
1587
1588
1589 static int __init mds_init(void)
1590 {
1591
1592         mds_file_cache = kmem_cache_create("ll_mds_file_data",
1593                                            sizeof(struct mds_file_data),
1594                                            0, 0, NULL, NULL);
1595         if (mds_file_cache == NULL)
1596                 return -ENOMEM;
1597
1598         class_register_type(&mds_obd_ops, status_class_var, LUSTRE_MDS_NAME);
1599         class_register_type(&mdt_obd_ops, 0, LUSTRE_MDT_NAME);
1600         ldlm_register_intent(ldlm_intent_policy);
1601         return 0;
1602
1603 }
1604
1605 static void __exit mds_exit(void)
1606 {
1607
1608
1609         ldlm_unregister_intent();
1610         class_unregister_type(LUSTRE_MDS_NAME);
1611         class_unregister_type(LUSTRE_MDT_NAME);
1612         if (kmem_cache_destroy(mds_file_cache))
1613                 CERROR("couldn't free MDS file cache\n");
1614
1615 }
1616
1617 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1618 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1619 MODULE_LICENSE("GPL");
1620
1621 module_init(mds_init);
1622 module_exit(mds_exit);